A nonprofit research lab studying where AI safety and security methods break, from hidden model behaviors and backdoors to benchmarks and evaluation.